def propose(self,trace,scaffold):
self.trace = trace
self.scaffold = scaffold
if not registerVariationalLKernels(trace,scaffold):
self.delegate = MHOperator()
return self.delegate.propose(trace,scaffold)
_,self.rhoDB = detachAndExtract(trace,scaffold.border[0],scaffold)
assertTorus(scaffold)
for _ in range(self.numIters):
gradients = {}
gain = regenAndAttach(trace,scaffold.border[0],scaffold,False,OmegaDB(),gradients)
detachAndExtract(trace,scaffold.border[0],scaffold)
assertTorus(scaffold)
for node,lkernel in scaffold.lkernels.iteritems():
if isinstance(lkernel,VariationalLKernel):
assert node in gradients
lkernel.updateParameters(gradients[node],gain,self.stepSize)
rhoWeight = regenAndAttach(trace,scaffold.border[0],scaffold,True,self.rhoDB,{})
detachAndExtract(trace,scaffold.border[0],scaffold)
assertTorus(scaffold)
xiWeight = regenAndAttach(trace,scaffold.border[0],scaffold,False,OmegaDB(),{})
return trace,xiWeight - rhoWeight
def propose(self,trace,scaffold):
from particle import Particle
self.trace = trace
self.scaffold = scaffold
assertTrace(self.trace,self.scaffold)
#print map(len, scaffold.border)
self.T = len(self.scaffold.border)
T = self.T
P = self.P
# assert T == 1 # TODO temporary
rhoDBs = [None for t in range(T)]
rhoWeights = [None for t in range(T)]
for t in reversed(range(T)):
rhoWeights[t],rhoDBs[t] = detachAndExtract(trace,scaffold.border[t],scaffold)
assertTorus(scaffold)
particles = [Particle(trace) for p in range(P+1)]
self.particles = particles
particleWeights = [None for p in range(P+1)]
# Simulate and calculate initial xiWeights
for p in range(P):
particleWeights[p] = regenAndAttach(particles[p],scaffold.border[0],scaffold,False,OmegaDB(),{})
particleWeights[P] = regenAndAttach(particles[P],scaffold.border[0],scaffold,True,rhoDBs[0],{})
assert_almost_equal(particleWeights[P],rhoWeights[0])
# for every time step,
for t in range(1,T):
newParticles = [None for p in range(P+1)]
newParticleWeights = [None for p in range(P+1)]
# Sample new particle and propagate
for p in range(P):
parent = sampleLogCategorical(particleWeights)
newParticles[p] = Particle(particles[parent])
newParticleWeights[p] = regenAndAttach(newParticles[p],self.scaffold.border[t],self.scaffold,False,OmegaDB(),{})
newParticles[P] = Particle(particles[P])
newParticleWeights[P] = regenAndAttach(newParticles[P],self.scaffold.border[t],self.scaffold,True,rhoDBs[t],{})
assert_almost_equal(newParticleWeights[P],rhoWeights[t])
particles = newParticles
particleWeights = newParticleWeights
# Now sample a NEW particle in proportion to its weight
finalIndex = sampleLogCategorical(particleWeights[0:-1])
assert finalIndex < P
self.finalIndex = finalIndex
self.particles = particles
return particles[finalIndex],self._compute_alpha(particleWeights, finalIndex)
def makeConsistent(self,trace,indexer):
# Go through every local child and do extra and regen.
# This is to be called at the end of a number of transitions.
if not hasattr(self, "global_scaffold"):
self.global_scaffold = indexer.sampleGlobalIndex(trace)
for local_child in self.global_scaffold.local_children:
local_scaffold = indexer.sampleLocalIndex(trace,local_child)
_,local_rhoDB = detachAndExtract(trace, local_scaffold.border[0], local_scaffold)
regenAndAttach(trace,local_scaffold.border[0],local_scaffold,False,local_rhoDB,{})
def reject(self):
# TODO This is the same as MHOperator reject except for the
# delegation thing -- abstract
if self.delegate is None:
detachAndExtract(self.trace,self.scaffold.border[0],self.scaffold)
assertTorus(self.scaffold)
regenAndAttach(self.trace,self.scaffold.border[0],self.scaffold,True,self.rhoDB,{})
else:
self.delegate.reject()
def propose(self,trace,scaffold):
self.trace = trace
self.scaffold = scaffold
assertTrace(self.trace,self.scaffold)
self.T = len(self.scaffold.border)
T = self.T
P = self.P
rhoWeights = [None for t in range(T)]
omegaDBs = [[None for p in range(P+1)] for t in range(T)]
ancestorIndices = [[None for p in range(P)] + [P] for t in range(T)]
self.omegaDBs = omegaDBs
self.ancestorIndices = ancestorIndices
for t in reversed(range(T)):
(rhoWeights[t],omegaDBs[t][P]) = detachAndExtract(trace,scaffold.border[t],scaffold)
assertTorus(scaffold)
xiWeights = [None for p in range(P)]
# Simulate and calculate initial xiWeights
for p in range(P):
regenAndAttach(trace,scaffold.border[0],scaffold,False,OmegaDB(),{})
(xiWeights[p],omegaDBs[0][p]) = detachAndExtract(trace,scaffold.border[0],scaffold)
# for every time step,
for t in range(1,T):
newWeights = [None for p in range(P)]
# Sample new particle and propagate
for p in range(P):
extendedWeights = xiWeights + [rhoWeights[t-1]]
ancestorIndices[t][p] = sampleLogCategorical(extendedWeights)
path = constructAncestorPath(ancestorIndices,t,p)
restoreAncestorPath(trace,self.scaffold.border,self.scaffold,omegaDBs,t,path)
regenAndAttach(trace,self.scaffold.border[t],self.scaffold,False,OmegaDB(),{})
(newWeights[p],omegaDBs[t][p]) = detachAndExtract(trace,self.scaffold.border[t],self.scaffold)
detachRest(trace,self.scaffold.border,self.scaffold,t)
xiWeights = newWeights
# Now sample a NEW particle in proportion to its weight
finalIndex = sampleLogCategorical(xiWeights)
path = constructAncestorPath(ancestorIndices,T-1,finalIndex) + [finalIndex]
assert len(path) == T
restoreAncestorPath(trace,self.scaffold.border,self.scaffold,omegaDBs,T,path)
assertTrace(self.trace,self.scaffold)
return trace,self._compute_alpha(rhoWeights[T-1], xiWeights, finalIndex)
def evalOneLocalSection(self, trace, local_scaffold, compute_gradient = False):
assert(self.global_scaffold.globalBorder is not None)
# Detach and extract
_,local_rhoDB = detachAndExtract(trace, local_scaffold.border[0], local_scaffold, compute_gradient)
# Regen and attach with the old value
proposed_value = trace.valueAt(self.global_scaffold.globalBorder)
trace.setValueAt(self.global_scaffold.globalBorder, self.global_rhoDB.getValue(self.global_scaffold.globalBorder))
regenAndAttach(trace,local_scaffold.border[0],local_scaffold,False,local_rhoDB,{})
# Detach and extract
rhoWeight,local_rhoDB = detachAndExtract(trace, local_scaffold.border[0], local_scaffold, compute_gradient)
# Regen and attach with the new value
trace.setValueAt(self.global_scaffold.globalBorder, proposed_value)
xiWeight = regenAndAttach(trace,local_scaffold.border[0],local_scaffold,False,local_rhoDB,{})
return xiWeight - rhoWeight
def propose(self,trace,scaffold):
from particle import Particle
assertTrace(trace,scaffold)
pnodes = scaffold.getPrincipalNodes()
currentValues = getCurrentValues(trace,pnodes)
allSetsOfValues = getCartesianProductOfEnumeratedValues(trace,pnodes)
registerDeterministicLKernels(trace,scaffold,pnodes,currentValues)
detachAndExtract(trace,scaffold.border[0],scaffold)
assertTorus(scaffold)
xiWeights = []
xiParticles = []
for p in range(len(allSetsOfValues)):
newValues = allSetsOfValues[p]
xiParticle = Particle(trace)
assertTorus(scaffold)
registerDeterministicLKernels(trace,scaffold,pnodes,newValues)
xiParticles.append(xiParticle)
xiWeights.append(regenAndAttach(xiParticle,scaffold.border[0],scaffold,False,OmegaDB(),{}))
# Now sample a NEW particle in proportion to its weight
finalIndex = sampleLogCategorical(xiWeights)
self.finalParticle = xiParticles[finalIndex]
return self.finalParticle,0
def propose(self, trace, scaffold):
self.prepare(trace, scaffold)
logBound = computeRejectionBound(trace, scaffold, scaffold.border[0])
accept = False
while not accept:
xiWeight = regenAndAttach(trace, scaffold.border[0], scaffold, False, self.rhoDB, {})
accept = random.random() < math.exp(xiWeight - logBound)
if not accept:
detachAndExtract(trace, scaffold.border[0], scaffold)
return trace, 0
def fixed_regen(self, values):
# Ensure repeatability of randomness
cur_pyr_state = random.getstate()
cur_numpyr_state = npr.get_state()
try:
random.setstate(self.pyr_state)
npr.set_state(self.numpyr_state)
registerDeterministicLKernels(self.trace, self.scaffold, self.pnodes, values)
answer = regenAndAttach(self.trace, self.scaffold.border[0], self.scaffold, False, OmegaDB(), {})
finally:
random.setstate(cur_pyr_state)
npr.set_state(cur_numpyr_state)
return answer
def propose(self, trace, global_scaffold):
rhoWeight = self.prepare(trace, global_scaffold)
xiWeight = regenAndAttach(trace,global_scaffold.border[0],global_scaffold,False,self.global_rhoDB,{})
return trace, xiWeight - rhoWeight
def reject(self):
detachAndExtract(self.trace,self.scaffold.border[0],self.scaffold)
assertTorus(self.scaffold)
regenAndAttach(self.trace,self.scaffold.border[0],self.scaffold,True,self.rhoDB,{})
def reject(self):
# Only restore the global section.
detachAndExtract(self.trace,self.global_scaffold.border[0],self.global_scaffold)
assertTorus(self.global_scaffold)
regenAndAttach(self.trace,self.global_scaffold.border[0],self.global_scaffold,True,self.global_rhoDB,{})
def propose(self, trace, scaffold):
program = trace.proposal_programs[self.program_name]
pnodes = [node for tar in program.tlist for node in trace.getNodesInBlock(tar[0],tar[1])]
cnodes = [node for cond in program.clist for node in trace.scopes[cond[0]][cond[1]]]
# drawScaffoldKernel(trace,scaffold,pnodes,cnodes,program.tlist,program.clist)
try:
assert len(pnodes) == program.n_target
except:
raise(Exception('Expect to have one and only one random node in a block. Check if 1) multiple nodes are defined in a target block; 2) some target nodes are observed (and thus not random).'))
old_target = [node.value.number for node in pnodes]
conditioned = [node.value.number for node in cnodes]
# print conditioned
registerDeterministicLKernels(trace,scaffold,pnodes,map(lambda x:VentureNumber(x),old_target))
rhoWeight = self.prepare(trace, scaffold)
# print "CustomMHOperator rhoWeight: ", rhoWeight, "old_target: ", old_target
# print rhoWeight
# TODO: get conditioned values
# TODO: invoke program, get target values and qratio from proposal
# print 'cond:', conditioned
# print 'old tar:', old_target
# (new_target, qratio) = program.propose(conditioned,old_target)
conditioned_labels, latent_labels, target_labels = [], [], []
conditioned_src = program.gen_conditioned_src(conditioned)
if conditioned_src:
(conditioned_labels, conditioned_strings, _) = execute_and_record(program.ripl, conditioned_src, "conditioned")
latent_src = program.gen_latent_src(conditioned)
if latent_src:
(latent_labels, latent_strings, _) = execute_and_record(program.ripl, latent_src, "latent")
target_src = program.gen_target_src(conditioned)
if target_src:
(target_labels, target_strings, new_target) = execute_and_record(program.ripl, target_src, "target")
else:
raise(Exception("No target_src found. Procedure 'gen_target_src' must be specified."))
for label in reversed(target_labels):
program.ripl.forget(label)
if self.method == 'assumed_gibbs':
qratio = 1e10
else:
old_logscores, new_logscores = [], []
if not latent_src: self.mc_samples = 1
for i in range(self.mc_samples):
predict_to_observe(program.ripl, target_strings, new_target)
new_logscores.append(sum_directive_logscore(program.ripl, target_labels))
for label in reversed(target_labels):
program.ripl.forget(label)
predict_to_observe(program.ripl, target_strings, old_target)
old_logscores.append(sum_directive_logscore(program.ripl, target_labels))
for label in reversed(target_labels):
program.ripl.forget(label)
for label in reversed(latent_labels):
program.ripl.forget(label)
execute_and_record(program.ripl, latent_src, "latent")
old_logscore = logaddexp(old_logscores)
new_logscore = logaddexp(new_logscores)
qratio = old_logscore - new_logscore
# print "CustomMHOperator old_logscores: ", old_logscores, "CustomMHOperator new_logscores: ", new_logscores
for label in reversed(latent_labels):
program.ripl.forget(label)
for label in reversed(conditioned_labels):
program.ripl.forget(label)
# print map(lambda x:VentureNumber(x),new_target)
# TODO: use DeterministicLKernel to fill in new target values
registerDeterministicLKernels(trace,scaffold,pnodes,map(lambda x:VentureNumber(x),new_target))
# TODO: get xiWeight
# print scaffold.border[0]
xiWeight = regenAndAttach(trace,scaffold.border[0],scaffold,False,OmegaDB(),{})
# print "CustomMHOperator xiWeight: ", xiWeight, "new_target: ", new_target
# print "CustomMHOperator qratio: ", qratio
# TODO: return (new_trace, xiWeight + qratio - rhoWeight)
# print '====================='
# print 'new_target', new_target
# print 'xiWeight:', xiWeight
# print 'qratio:', qratio
# print 'rhoWeight:', rhoWeight
return (trace, xiWeight + qratio - rhoWeight)
def restoreAncestorPath(trace,border,scaffold,omegaDBs,t,path):
for i in range(t):
selectedDB = omegaDBs[i][path[i]]
regenAndAttach(trace,border[i],scaffold,True,selectedDB,{})
